On Mon, Feb 22, 2010 at 12:09 PM, Leigh J. Halliwell
<[email protected]> wrote:
> I'm quite profficient in the standard linear algebra with linear and
> quadratic forms of random vectors.  So the variance of an (m x 1) random
> vector x is an (m x m) matrix whose ijth element is cov(x[i], x[j]), and
> Var(Ax) = A Var(x) T(A).
>
> But I'd like to take the variance of an (m x n) random matrix X.  This would
> be an (m x n x m x n) array (or better, an ((m x n) x (m x n)) matrix) whose
> (ij)(kl)th element is cov(x[ij], x[kl]).  And I'd like also to generalize
> that quardratic form.
>
> So I'm seeking to generalize matrices from {linear mappings from n-space to
> m-space} to {linear mappings from (n1 X n2)-space to (m1 x m2)-space}.  At
> the very least, I can ravel (,) the axes and do it as (m1*m2) x (n1*n2)
> matrices.  But I wonder if something more powerful exists.

I have to admit that I am still lost.  Someone like John Randall
might be able to help you, but I am a complete novice when
it comes to this kind of thing.

I can perform experiments:

   X=: ? 5 5$0
   Y=: ? 5 5 $0
   E=:(+/%#)@,
   E ((-E)X) * ((-E)Y)
_0.0339656
   E ((-E)X) +/ .* ((-E)Y)
0.00833635

But I do not have the background to assign any meaningful
concepts to these results.

That said, I do have a question that might (or might not)
help me:  What do you have in mind for the expected
values in the multi-dimensional case you have described?
(Are you working with any dependencies?)

-- 
Raul
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to